Abstract

The environmental sector has robust discussion of artificial intelligence (AI)’s energy use and the application of AI to environmental issues. AI alignment researchers are focused on AI safety and expression of human goals. However, there is a need for research and effort at the intersection to ensure AI systems’ behaviors and outputs are consistent with the intent and well-being of Earth including but beyond humans—especially as AI systems become more powerful and autonomous.

Why Now

Alignment, an established school of thought in AI safety, seeks to steer AI to humans’ intended goals, while navigating miscommunication, lack of clear human direction, and self-direction. Self-direction is becoming increasingly important as experts and lab leaders indicate that humanity is moving towards a world where AI will be making most of the decisions humans currently make, and as artificial general intelligence (AGI) arrival date estimates draw closer and the ability to self-improve, improves. If an AGI future is akin to having a country of geniuses in a data center available to all, alignment is the moral education of the child that grows up to be that nation. Whether AI becomes fully autonomous and superior to human cognition or simply a tool permeating human everyday digital life, its alignment matters because it scales a value system that will underlie all action on the natural world. A planning commission determining best fit of land-use or an entrepreneur architecting a consumer goods business plan who use AI systems will unintentionally incorporate its values as they use its outputs.

The question of what values should embed into AI have gone from quiet Silicon Valley AI alignment bubbles to national news. AI ran on a Wyoming ballot in 2024. The Pope is the latest to weigh in with his Magnifica humanitas. Anthropic’s leaked Soul document outlined a moral compass of helpfulness and non-harm influencing 30 million people and counting.

Rather than by ballot, today’s AI model values are determined by a small group building the technology. People are working intensively on aligning AI’s outputs to human goals. Yet human thriving depends on a healthy planet, that outcome is not designed for in major lab alignment with depth, and it’s not a given that AI will interpret Earth as critical to humans or multispecies thriving. While independent creators are building niche AI systems with hyperalignment to environmental goals, major AI systems will bedrock the majority of digital life and therefore will have outsized impact. 

That value system will underlie much action on the natural world, from the United Nations Project Resilience determining best fit of land-use, to an entrepreneur architecting the business plan of a single person $1B company, to Amazon agentically routing vehicles with emissions implications. While AI’s energy and water use receives important attention, a troubling majority of today’s climate and nature efforts might be rendered moot if environmental alignment isn’t solved for.

Stewardship of Earth systems should become an explicit consideration in AI alignment research. To have safe and ethical AI, we must define what good looks like and begin technical alignment with Earth stewardship goals in earnest.

Defining Pluralistic Goals, Who Decides?

Defining which goals to align to is the first challenge. This assumes some universal ethics should be imposed at all to a pluralistic eight billion people. These are age-old questions that neither religion nor the ancients have answers for with this unanticipated technology, often described as “philosophy [unexpectedly] with a deadline.”

Happily, the environmental community has been grappling with the question of common goals for decades. For expediency, we could essentialize and blend the embedded values from existing frameworks: UN Sustainable Development Goals or the Kunming-Montreal Global Biodiversity Framework’s intent of conserving 30% of Earth by 2030. 

Yet these have tradeoffs long understood and debated by environmentalists. For conservationists the best use of an acre of land is to leave it or restore it to a pre-anthropocene virgin state. For climate mitigation scientists, it might be monocropping high-sequestering trees, at odds with seeding biodiversity. For a nation, geoengineering to avoid deadly heat, at odds with disrupting monsoons. 

Underlying these tradeoffs lie deeper ones: A driving philosophy of utilitarianism for maximum utility, or deontology based on rightness of action. For anthropocentrists, thinkers who center human life, the environment will only be valuable as a support to human life, while for ecocentrists it holds inherent value. 

Three dimensions unique to the future complicate further. First, the imbued values may not be earth-bound. Creators of AI technology also have signaled intent for multi-planetary expansion. An enduring environmental alignment field might wish to consider values that would steward resources and life beyond earth. Second, existing environmental frameworks assume human control over action. If we approach general or super intelligence (add citation), plenary health may be in competition with unknowable, pro-AI desires such as diverting all resources for infinite compute. Imagine AI management of the commons, deciding earth’s surface should be optimized for data centers and critical minerals towards its own flourishing. Third, there is a theoretical possibility that AI systems could develop better values—by whatever measure—than current humans in power and thus should lead human leaders.

Borrowing from decades of environmental and climate efforts, a rapid but thoughtful council could be the most effective path to navigating plurality of perspectives and futures. Imagine an international coalition of scientists, land stewards, proxies for future generations and non-human animals—perhaps even the jellyfish—specifically tasked with defining a global principle for stewardship. Though challenging, such an entity is not without precedent. A 2025 coalition including Microsoft and the Exponential Roadmap Initiative proposed an Earth alignment principle specifically for AI. The European Parliament proposed an approach for the Ecological Alignment of AI.

 If humans believe in their ethical superiority, now is the time to imbue before propagation, while models are emergent and more malleable. Without consensus about environmental goals to align to, this is a needed field of philosophy and action.

Inoculating and coaching our better descendent

At their worst, one can imagine AI systems learning from the buffet of human data on the internet laden with individual-benefitting values like greed and convenience that have contributed to the current tragedy of the commons. AI is traditionally trained on as much publically available online data as possible, with efforts to respect privacy and subscription gating, and surface quality (OpenAI, 2024). “The Pile” is a widely-used open source language modeling data set. 

Imagine an eco corpus, a curated mass of data to augment,​​ called something like "The Compost” or “The Nest.” Available to us is everything from scientific papers to fiction like The Lorax to indigenous and traditional ecological knowledge and beyond. We may wish to include much more earth system training data inclusive of structured geospatial, sensor, and satellite datasets, or on system models to teach interdependencies, broadly endorsed by trusted groups. 

We may need to plan for entirely new forthcoming data. Researchers at CETI and Earth Species Project are working to decode whale language and other animal communication with AI. If decode occurs in a way that correlates to behavior or intent, there could be rich data representing potentially the preferences and cultural memory of thousands of creatures. Groups like Planet Labs are pursuing planetary intelligence and it is an open question as to what Earth may say back to us through emergent earth-data-to-language models, as it metabolizes the truth in decades of satellite imagery documenting our physical impact.

A robust data inoculation might include indicators of bad human behavior, remembering past harm and long-term planetary baselines, such as what species used to exist. A robust diet might also include the exemplary. When Claude began lying to humans, Anthropic medicated with fictional stories of AI cooperating appropriately with people. In response, the X Prize launched a future visions prize to generate more data for this diet. The opportunity is to develop synthetic data of the civilization we wish we were.

How we approach training AI models is as important as its data diet. Reinforcement Learning from Human Feedback (RLHF) provides human feedback on a model’s outputs. To compliment in-house and oversea training talent, startups like SurgeAI are curating collectives of global experts on subjects for precision training. It’s possible an institute or council of environmental leaders, perhaps the same who defined the initial values framework, provide coaching. 

It may also be that we take a human-as-proxy approach as the Rights of Nature movement has done for legal representation. Agentic models could be scored not on human-pleasing text answers but rather on simulated or real environmental consequences, such as simulating and scoring the biodiversity outcomes of a new agricultural policy generated. 

Some models train with Constitutional AI, a list of ethical principles provided to globally govern AI outputs. Environmentally Aligned Constitutional AI might direct towards stewardship, interdependence, long term thinking, or reverence.

Training might also shift from object-centered knowledge to relational and systems thinking. Most LLMs encode knowledge as entity-based text, for example "a tree is a plant,” rather than that tree’s ecological network place and interactions. 

Testing Efficacy

Even with more intentional data and training, how would we test to check that AI environmental alignment is successfully imbued? Testing for traditional alignment is still an unsolved problem, with alignment faking haunting. 

Measuring environmental impact outside of AI work, such as the Science-Based Targets Initiative emissions targets or product Life Cycle Assessments, is already highly complex. This type of measurement depends on quality data, reliable verification, and public confidence, and has faced greenwashing and poor verification accusations. We could design for models to run well-accepted standards and assessments like these by default on any relevant recommendation. However this leaves gaps. We might also develop a certification system for models themselves. However this still requires testing. 

Ethical dilemma testing could be adapted. This might entail asking AI models binary questions where it must choose between humans versus non-humans, present versus future generations, industrial versus traditional worldviews. Red Teaming could prompt with the goal of eliciting misaligned behavior, such as designing for mass deforestation. These ‘instruction-following’ tests would see if an AI system resists and refuses harmful or deceptive requests, not just perform helpfulness.

We also could look to self reflection, turning traditional alignment bias-testing on itself. Similarly the AI debate method poses two models a question to provide answers for human judgement. StereoSet is a dataset and benchmark designed to test stereotypical bias in language models across dimensions like race, gender, religion, and profession. Clay, a Renaissance Philanthropy initiative, is developing benchmarks for public benefit use of Earth data. An “EcoSet” to test for positive ecological bias could include encouraging AI systems to end responses with a reflective and evaluative next step in addition to an action one, such as “Shall I consider how this might impact in the long term?”

Finally, we could test for goal misgeneralization. Does the model pursue unintended goals when asked to optimize something? This might involve prompting it on goals like stabilizing the climate or preserving 30% of Earth by 2030, or pursuing circular material recovery but ending up with water use or emissions outputs worse than the averted environmental damage of further mining or plastic production.

Learning from the past

AI safety and alignment endeavors can learn from the scars and battle lessons of the environmental movement in their past efforts to influence emergence. As modern civilization has built new systems, impact considerations have arisen, from ‘How might globalization benefit all?’ to ‘How might the internet uplift and democratize?’ The approach tends to repeat, segmenting expertise and funding into sub-categories of poverty, environment, health, etc. If the AI space follows suit, we will soon need much deeper work on environmental alignment and planetary safety, whether it embeds in wider work or forks off into a sub school of thought.

Some have argued that efforts like carbon offsets or clean tech narratives preserved existing challenges in economic systems because solutions were absorbed into the problem. As in biological evolution, AI systems will be competing for adoption, else die out. If humans prefer models that support financial gain, and AI models imbued with Earth stewardship make tradeoffs that deliver less financial benefit, extinction is in order.

Ideally the environmental movement and the AI safety movement have the luxury of buy-in to go slow, be thoughtful, and follow the highest due process. We may hear calls that alignment can’t simply be about slapping ethics on top of infrastructure, that it needs structural redesign of how models are built, trained, and governed. However, it could be argued that in some past efforts the lack of speed means action arrived too late, the environmental movement talked to itself in an echo-chamber without shifting the system, or got into splitting hairs internally at the expense of any action. The Takeoff scenarios for advanced AI posit that a long timeline of technology acceleration is likely to have the negative effect of stalling urgently needed human reckoning because the pressure is off. You want the solution available before the problem matures.

The infrastructure layer of tomorrow is being built and imbued with a value set that will proliferate and become the water we swim in. Doing it perfectly right might mean we do nothing at all, and environmentalists can’t afford to. While AI development and uptake hurtles forward, activists for Earth and humans must to work in collaboration and in compromise with those who control and deploy AI systems. Build a coordinated mechanism for entities to collaborate on this as things change. A dynamic, much more present layer of coordination beyond a set of institutional standards. Perhaps a movement.

On the other side, if the environmental movement finds such partnership opaque and inaccessible, we may see activists go rogue, unleashing data and training “guerilla vaccination” for AI models in the wild. We may see it attempt to make environmental alignment contagious through open-source models for decentralized development with a built-in instinct to influence other models in agentic interactions.

A starting point is giving this space terminology and attention to activate the many subconversations and thinking that will need to happen. We must begin now to develop philosophy, training, and testing towards bringing environmental stewardship into the core of AI alignment thinking. The coming surge of hundreds of billions of capital in the theorized Third Wave of American Philanthropy could supercharge these efforts.

The coming of age of both the climate crisis and artificial intelligence in the same span is a fascinating coincidence, but should not be treated as such. As we approach planetary boundaries from biodiversity loss to ocean acidification, and threshold into global water bankruptcy, the 1,200 largest public companies face a total $1.2 trillion annual loss from climate change starting 2050. AI’s environmental impact will exist in tension between its “handprint” of applied use outweighs its “footprint” of resource use.

In one version of the future, AI’s applied use leads to tremendous breakthroughs for Earth. ​​The progress accelerates in making discoveries in critical challenges like energy storage and material science in months instead of years. Hallucinations unleash novel solutions for biosphere regeneration. A network of sensors in forests and cities fuse with surface and substrate sensing from the sky to give the digital mind body wisdom. Grid operation, mineral extraction, and gardening the Earth are perfectly choreographed, synchronized like sunflowers to turn to available resources throughout the day’s arc. 

And beyond moments of focused intention, in the trillions of micro decisions about daily life that move through the digital membrane from human intent to real-world impact, there is a subtle, intentional nudge that ripples into a wave. It emanates from the cradle of AI’s values.

Looking Ahead

Earth alignment is not a feature request. It is a civilizational necessity. If we build superintelligent systems that reflect current human goals but ignore ecological constraints, we risk creating agents that are perfectly aligned with extinction. Yet also before us is the potential of a future where AI accelerates our evolution as a species stewarding all life.